Anthropic Locks in $517 Billion in Compute Ahead of IPO, Still Trails OpenAI
nashnova research
Anthropic signed $517 billion in compute deals over 11 months — nearly triple its disclosed $180 billion plan — yet OpenAI's $750 billion spending target means the capacity gap remains wide as Anthropic's IPO prospectus nears release.
Where does the $517 billion go?
The total spans three buckets: cloud-capacity leases, chip purchases, and data-center leases, all signed since last October.
Anthropic has locked in at least 14.8 GW of compute capacity, plus 1–2 GW secured earlier — bringing the total close to 17 GW.
This means → in under a year, Anthropic committed roughly three times what it told investors it would spend through 2029.
Who is supplying the power?
Early backers Amazon and Google remain the largest providers, delivering 11 GW combined at a cost exceeding $300 billion over roughly ten years.
New suppliers have been poached from OpenAI's orbit: Microsoft Azure supplies about 1 GW for at least $30 billion; Elon Musk's SpaceX leases AI servers at $1.25 billion per month through May 2029, totaling up to $45 billion — though either side can cancel with 90 days' notice.
Startups Lambda and UK-based Nscale won $80 billion in cloud contracts, locking in at least 460 MW for six years.
In plain terms = Anthropic is spreading its bets across as many suppliers as possible — and recruiting directly from its rival's supply chain.
What about chips and owned data centers?
On the chip side, Anthropic signed a multi-GW deal in April for Google's tensor processing units (TPUs — AI chips designed by Google, built by Broadcom), then added a 2 GW AMD chip order in July.
For owned infrastructure, a $50 billion joint investment with cloud startup Fluidstack is building data centers in Texas and New York; some facilities are already operational.
A lease with former Bitcoin miner TeraWulf adds a Kentucky AI data center that will eventually deliver 401 MW.
Revenue leads — so why does compute still lag OpenAI?
Anthropic's annualized revenue now exceeds $65 billion, surpassing OpenAI's $40 billion-plus as of July — it is winning on the income side.
But OpenAI plans to control 30 GW by 2030 and last month projected spending of roughly $750 billion by then, up from an earlier $665 billion estimate.
This means → Anthropic is making money faster, yet its rate of stockpiling compute still trails OpenAI's expansion pace — the arms-race gap has not narrowed.
What should investors look for in the IPO prospectus?
Signed capacity may not all be used, and Anthropic likely has additional undisclosed agreements — the published figures are a floor, not a ceiling.
Data-center construction faces site selection, grid-connection, and permitting hurdles; there is real slippage risk between contract and delivery.
In plain terms = what investors need the prospectus to answer is: of that $517 billion, how much becomes running compute and how much stays a placeholder contract — that answer will directly set the market's pricing of margins and customer retention.
市场有风险，内容仅供研究参考，不构成投资建议。